getsentry / getsentry/sentry-javascript
Node: process never exits since 10.72 — beforeExit session send + client-report flush feed each other (skipOpenTelemetrySetup: true)
- Lenguaje dominante
- TypeScript
- Estrellas
- 8.7k
- Forks
- 1.8k
- Merge medio
- 1 d 17 h
- PR fusionados (30 d)
- 523
Descripción
### Is there an existing issue for this?
- [x] I have checked for existing issues https://github.com/getsentry/sentry-javascript/issues
- [x] I have reviewed the documentation https://docs.sentry.io/
- [x] I am using the latest SDK release https://github.com/getsentry/sentry-javascript/releases
### How do you use Sentry?
Sentry Saas (sentry.io)
### Which SDK are you using?
@sentry/node
### SDK Version
10.73.0 (also 10.72.0; 10.54.0 is fine)
### Framework Version
Node 22.22.2
### Link to Sentry event
_No response_
### Reproduction Example/SDK Setup
```js
// repro.mjs — run with a DSN pointing at any HTTP server that answers 200, e.g. a stub on localhost
import * as Sentry from "@sentry/node"
Sentry.init({
dsn: process.env.SENTRY_DSN,
release: "1.0.0", // any release, so the process session is sendable
skipOpenTelemetrySetup: true, // we run our own OTel pipeline elsewhere
tracesSampleRate: undefined, // tracing off
})
console.log("done; letting the process exit naturally")
```
Stub receiver used for the counts below:
```js
import http from "node:http"
let n = 0
http.createServer((req, res) => { let b = ""; req.on("data", c => b += c); req.on("end", () => {
n++; console.log(n, req.url, [...b.matchAll(/"type":"([a-z_]+)"/g)].map(m => m[1]).join(","), b.slice(-160)); res.end("{}") }) }).listen(9999)
```
### Steps to Reproduce
1. `node stub.mjs`
2. `SENTRY_DSN=http://k@localhost:9999/1 node repro.mjs`
### Expected Result
The process sends the session (and at most one client report) and exits.
### Actual Result
The process never exits. The stub receives one `session` envelope and then an endless stream of `client_report` envelopes, each containing:
```json
{"discarded_events":[{"reason":"no_parent_span","category":"span","quantity":1}]}
```
CPU sits at 60–100% against a local stub; against a real DSN the loop is network-bound but the process is still immortal. We hit this in a Cloud Run job running database migrations, which hung until its 600 s task timeout on every deploy after upgrading from 10.54.
What appears to happen:
1. #23731 (10.72) makes `processSessionIntegration` end the session on `beforeExit` when its status is `ok`, so a healthy process now sends a `session` envelope at exit (previously the condition was inverted and healthy processes sent nothing).
2. That transport request is instrumented by the http integration. With no OTel setup and tracing off, its span is dropped and `recordDroppedEvent("no_parent_span", "span")` records an outcome.
3. The client-report `beforeExit` listener flushes the new outcome → another instrumented request → another `no_parent_span` outcome → `beforeExit` again → …
Confirmed by elimination:
| init options | result |
|---|---|
| `{ dsn, release }` | 1 session, exits |
| `{ dsn, release, skipOpenTelemetrySetup: true }` | 1 session + infinite client reports, never exits |
| `{ dsn, release, skipOpenTelemetrySetup: true, sendClientReports: false }` | 1 session, exits |
| same as row 2 on 10.54.0 | nothing sent, exits |
`await Sentry.close()` before exiting also avoids it, but the default behaviour of a process that simply finishes should not be to hang.
### Additional Context
Suggested fixes: don't record a `no_parent_span` outcome for the SDK's own transport requests, and/or don't re-arm the client-report flush from within a `beforeExit` flush (e.g. only flush outcomes that existed before the current `beforeExit` tick).
### Priority
Medium
Guía de contribución
Evaluación
Este issue todavía no se ha evaluado.